Nativ Packages MLX-VLM, Local APIs, and Performance Telemetry into an Open-Source macOS Inference Workspace

Running open models on Apple Silicon typically requires users to manage Python environments, MLX services, model caches, API adapters, and performance monitoring themselves. Nativ v0.0.1 attempts to consolidate these components into a single macOS app: its frontend is built with SwiftUI, `NativServerKit` manages a relocatable Python environment and the server lifecycle, and the underlying models run through `mlx-vlm` and MLX using unified memory. The project attracted attention in a July 26 roundup from the local AI community, and its GitHub page currently shows approximately 925 stars.

Nativ scans the Hugging Face cache and uses the device’s available memory to indicate whether a model is suitable for loading. Its interface also displays time to first token, decoding speed, token usage, GPU utilization, memory usage, and swap pressure. Of greater engineering value, it provides more than a chat UI: on localhost, it exposes OpenAI-compatible Chat Completions, Responses, Models, Images, and Audio routes, as well as Anthropic Messages and token-counting endpoints. This allows Codex, Claude Code, Pi, Hermes, and OpenCode to connect to models running on a Mac without major client-side modifications. Advanced settings also cover KV cache quantization, prefix caching, speculative decoding, structured outputs, and thinking budgets.

However, the initial release still has clear limitations. It supports only Apple Silicon and macOS 26 or later, while model loading is constrained by unified memory capacity and `mlx-vlm` compatibility. Dedicated speech and image-generation models remain planned for future releases. Although inference can run offline after a model has been downloaded, an internet connection is still required for the initial model download and dependency setup. The project also provides no standardized benchmarks for throughput, power consumption, long-term stability, or full OpenAI and Anthropic API compatibility rates. Adopters should first test it with real coding agents, tool calls, and long-context workloads instead of treating “compatible endpoints” as a drop-in replacement for cloud services.